Embracing the Future

GenAI’s Developing Role in Finance

A recent J.P. Morgan webinar – The Paradigm shift in financial services caused by Generative AI – lined up some key witnesses to explore how Generative AI, and large language model-based algorithms, could shape the future of European finance. TMI listened attentively.

There’s no doubt that Generative AI (GenAI) has the potential to make massive changes to the way the financial sector operates. Whether it does, or is allowed to, is still a moot point – after all, the technology is new, and the rules of engagement are yet to be codified in any meaningful way.

Its nascent capability is evident though. In the payments space in particular, GenAI can automate fraud detection, assess transaction risk, and optimise sanctions screening and other processes, for example. GenAI-driven chatbots and virtual assistants are arguably improving the customer service experience for many consumers and businesses. And predictive analytics is now capable of enhancing cash flow management, even ensuring regulatory compliance, as instant payment adoption increases in many regions.

It seems like a gift to the industry and its retail and commercial clients. Is it?

Sam Beck, Managing Director Head of Applied AI, Payments Operations, J.P. Morgan Payments, brought the webinar’s GenAI discussion into the wider arena of large language models (LLMs). LLMs are deep learning algorithms that can recognise, summarise, translate, predict, and generate content. They use a far wider range of architectures and applications than GenAI solutions such as ChatGPT. The technology, he revealed, is making significant changes to the way that the bank works, but admitted it has been a steep learning curve.

“When we first started, we embraced neural networks, and the concept of language models in general, but we had to start from scratch,” Beck explained. “We were training on our own data, creating our own architectures, and basically custom-fitting a model to the business challenges we were trying to solve. What’s interesting now is that we are approaching a situation where we have some starting-point models that are actually quite good at solving some of those business challenges.”

So far, Beck has accepted that there has not been a situation where the models can simply be installed and left to run. But he feels there has been significantly more success by starting with a pre-trained model, which is fuelling GenAI, and fine-tuning that to J.P. Morgan Payments’ use cases. “That makes the models more accurate, shortening the time to delivery into those business use cases – and that is having a big impact.”

Filling in the gaps

A key element of the generative element of AI is its capacity to create new content, noted Sudhir Jha, Executive Vice President, Mastercard. While mostly being used for text, he said the technology can be used for many more scenarios. The vast global Mastercard transactional data set can, for instance, be deployed to predict transaction success rates, and therefore the likelihood of anomalies occurring. “The more data that can be applied to anomaly detection, the more the accuracy and utility improves.”

Before LLM, Jha noted that nobody had tried to work with such a huge volume of data. But by using GenAI, it has enabled “more human-like conversations”. This has created “tremendous benefits” across the wider fintech world, where he believes many back- and front-office processes can be improved by “conversational interfaces” for both customer and internal communications.

“The way to think about GenAI is that it can help a business figure out all of its data assets – including unstructured text – and make them available to the right people, at the right time, in a more familiar and friendly format,” commented Jha. He explained that by doing so, it facilitates an enhanced understanding of anomalies and potential risk elements such as fraud, even intelligently filling in the gaps where data points are missing to create likely outcomes.

With the broad client base of a major global consultancy, Anu Widyalankara, Payments Strategy and Technology Lead, EY, concurred with these thoughts. She added that progress in this space is rapid, with the pool of LLMs having expanded “some way beyond what was available even last year”. With some of the regulations around data access now in place, she noted, too, that it provides a robust landscape for new market entrants. And with GenAI able to be built using a low-code/no-code easy-coding approach, her view is that it is opening up the field for many more to become involved.

With the conversational feel of interactions on digital channels, Widyalankara believes that Gen-AI enhances the customer onboarding experience, as indeed it does ongoing account management. And with it also paving the way for dynamic risk management modules to work in the background of these interactions, relationship managers will be afforded the chance to offer services such as real-time credit decisioning for clients.

What’s more, Widyalankara sees potential for GenAI to advance end-to-end transactional capabilities. “From a payment being captured, right through to its settlement, there are many improvements possible,” she noted, drawing attention to payments investigations as one of the major pain points for clients.

One specific area where Widyalankara thinks GenAI will offer a high level of assistance is embedded finance and payments. The likes of Klarna, Adyen and Stripe are all currently vying for position with their embedded offerings in an expanding range of verticals. These solutions are able to drive end-to-end transactions, and as a business progressively creates and embeds code for these solutions, GenAI can be deployed, for example, to monitor customer behaviour, enabling services to be adjusted for a range of platforms and channels.

Risk control

As a risk management tool for banks, Beck said LLMs used in the area of sanctions screening can be treated as a “supervised learning problem”. This plays to the strength of ML and language models, particularly when handling payments files that contain some unstructured text.

“We’ve also seen that when training a model based on historical data, including human decisions, there tends to be a quicker – almost real-time – and more consistent output, with more ‘explainability’ than might be achievable by human decision-makers. This is because the model has greater capacity to record its decisions and offer the rationale behind them.”

ML, and LLM in particular, can have a huge role to play in reducing the risk profile of banks, and provide greater process efficiency and an enhanced customer experience. In risk management processes such as KYC, simple initial checks have not been an issue for ‘normal’ automation. However, when exceptions arise and deeper investigation is required, it becomes an automation challenge, both technically and in terms of the cost of such an exercise.

“The kind of questions asked would be much easier for the client to respond to if structured in a conversational manner, as with GenAI,” Jha suggested. He added that the report generation that often applies to risk management is another “low-hanging fruit” for GenAI models, and noted too that “there will be many more innovative use cases” to come.

Making a difference

Should these tools take off, eventually all banks and businesses will have the same power at their disposal. It means their attempts at differentiation will surely be challenged. “This is when it comes back to the customer experience,” stated Widyalankara. However, while she acknowledged that foundational data architectures are being constructed to support enterprise-wide data access, her assessment is that most users are “still miles away from achieving an LLM that works equally across the organisation”. Technological differentiation remains possible, at least for now.

Jha offered a more nuanced take. While he noted that AI in general has indeed for some time been used as a means of differentiation for its early adopters, it has done so by “giving somewhat unfair advantage to those with access to proprietary data”. For him, then, the technology is only part of the equation, and that data access advantage will hold true for GenAI too, “because only those with the know-how to use that data, in the context of their unique business circumstances, will be able to differentiate through it”.

Beck agreed. The differentiating advantage of older tools such as OCR was that its algorithms were developed by third parties over a long period, and then adopted by a limited set of clients as a “proprietary magic sauce”. Today, state-of-the-art data extraction tools are often open source and thus readily available to many more users. Being able to train these tools the organisation’s own vast but unique data set is therefore what will set it apart from another, not the technology alone.

Quashing increased fraud risks

As with all digital financial solutions, fraud is an ever-present danger, and real-time processing presents an opportunity for real-time deception. With the rapid growth of instant payments having coincided with the pandemic-propelled increase in appetite for online payments, it shifted a mechanism from being what Jha described as a “benign P2P cash substitute”, to being a system used to rapidly send much larger volumes and sums, with a significantly increased potential for fraud.

As such, banks have been forced to accelerate their response not only to the arrival of a multitude of competitive non-bank real-time payment providers (a competitive environment set up in Europe by PSD2), but they have also had to put in place effective processes to counter the resulting rise in real-time criminality. For Jha, AI-based real-time detection is the only possible solution. “I can’t see any other way for this to adapt and scale.”

Key to his belief is the expectation that GenAI will be able to provide much more synthetic fraud data which, he explained, can be used to train AI models over much shorter timescales. “Banks cannot afford to wait to build up the necessary transaction data from which AI can learn; they need to do it today, and GenAI is the only way to achieve that compression of time.”

For an organisation to be truly comfortable with its own level of risk exposure, a dynamic approach to risk management is now necessary, added Widyalankara. “With instant payments, it’s essential that pre-validation checks – using tools such as the UK’s Confirmation of Payee – are able to leverage the maximum capabilities available for fraud and financial crime prevention, and GenAI plays very well in that space.”

Getting it right

Humans have proven to be adaptable towards new technologies, Beck noted. Just as most who fly in planes have no real idea how they work, simply accepting that they do, so many will trust in the use of innovative technologies if they provide a new capability. “GenAI is no different. It answers questions quickly, it offers a level of personalisation, and people are increasingly comfortable with that experience,” he observed.

But GenAI is typically trained to give answers, not to question the veracity of its resources. A greater understanding and control of the data that it consumes (whether text or numerical), is thus necessary. For businesses, this is certainly the case before releasing generated content into the wider world (as a Chatbot service discussing a bank’s financial products, for example).

Aside from factual inaccuracies, Jha’s other major concerns around content drawn from multiple sources are copyright infringement and data bias. But he suggested that within the next year or so, the debate on GenAI data integrity will have sharpened considerably as professional users begin to understand the rules of AI engagement, acknowledging that its use includes a duty of care to both internal and external consumers of its output.

It helps that the regulatory discussion is now moving ahead, said Widyalankara. In Europe, the world’s first comprehensive AI law, which covers the use of GenAI, is moving rapidly towards EU-wide agreement. And in the US, the Biden administration’s October 2023 executive order demands AI is “safe, secure and trustworthy”. Orders such as these that (in the case of the US) are taken “for the sake of our security, economy, and society”, help ensure developers are mindful that they are no longer operating in an AI Wild West.

New world order

Will AI disrupt the finance world? Some are trying hard to push the boundaries. As part of its programme of investment in AI-based tools, Beck revealed that J.P. Morgan has been thinking around the use of retrieval-augmented generation (RAG) technology. This, he explained, is an AI framework that uses GenAI to improve LLM-generated query responses from internal data. It can do so by augmenting those responses with the most up-to-date and accurate external information. RAG uses a natural language interface which can be used in diverse applications, from policy and procedure setting to client services.

“One could, in theory, get closer and closer to the customer with some of these techniques,” noted Beck. Deploying RAG could aid customer service personnel to access a bank’s broad range of proprietary information “to help the customer feel like they are speaking to a super-empowered agent.”

Another area of research within J.P. Morgan is conversational analytics. Rather than use traditional dashboards to display set fields, a user could ask a question in a natural language and an ‘artificial agent’ with access to the client’s reference and transactional data, could provide a much broader set of views, for example an assessment of how best to respond to certain liquidity scenarios.

For Beck, it will be customer-facing and internal querying aspects that will be most disrupted by GenAI in the future, largely because it could all be offered at a scale and convenience unachievable by humans. However, he suggested that effective delivery will require development of a new banking skill set, with traditional departmental experts required to work alongside AI, training the artificial agents to act on their behalf.

GenAI has so far typically been deployed in use cases where AI has already been applied to solve issues, noted Jha. In the coming year, he said he would like to see “some brand-new use cases”. But he offered a vision even further into the future too, where GenAI could be used to create what he called machine-to-machine (M2M) payments.

Here, a corporate account owner’s artificial agent could, for example, use its access to internal and external data to negotiate with the bank’s own artificial agent to delay a scheduled payment because an expected receivable is late. In this instance, the working capital advantage is obvious.

Extend simple M2M payments out to securing financing, optimising cash and liquidity, managing FX, and risk mitigation, and suddenly “it could look like a very different future”, mused Beck. It opens up a whole new discussion on what treasury might look like in a few decades, and in his concluding thoughts, he urged all to “embrace the change”.